@arcote.tech/arc-ai 0.7.8 → 0.7.9

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@arcote.tech/arc-ai",
   "type": "module",
-  "version": "0.7.8",
+  "version": "0.7.9",
   "private": false,
   "description": "AI provider abstraction, completion tracking, and budget management for Arc framework",
   "main": "./src/index.ts",
@@ -10,8 +10,8 @@
     "type-check": "tsc --noEmit"
   },
   "peerDependencies": {
-    "@arcote.tech/arc": "^0.7.8",
-    "@arcote.tech/arc-auth": "^0.7.8",
+    "@arcote.tech/arc": "^0.7.9",
+    "@arcote.tech/arc-auth": "^0.7.9",
     "typescript": "^5.0.0"
   },
   "devDependencies": {

package/src/types.ts

@@ -119,6 +119,12 @@ export interface CompletionRequest {
   webSearch?: boolean;
   temperature?: number;
   maxTokens?: number;
+  /**
+   * Reasoning effort dla modeli z reasoning (gpt-5, o-series). "minimal"
+   * pomija reasoning step → szybki time-to-first-token. Domyślnie provider
+   * wybiera (gpt-5 = medium). Adaptery bez wsparcia ignorują pole.
+   */
+  reasoningEffort?: "minimal" | "low" | "medium" | "high";
 }
 
 export interface CompletionResult {
@@ -140,22 +146,34 @@ export interface CompletionResult {
 
 // ─── Streaming ───────────────────────────────────────────────────
 
-export type StreamEventType =
-  | "content_delta"
-  | "tool_call_start"
-  | "tool_call_delta"
-  | "tool_result"
-  | "usage_update"
-  | "done"
-  | "error";
+/**
+ * Provider emit'uje cztery fazy tool calla:
+ *   started → arguments_delta(*) → arguments_complete → (server-side execute)
+ *
+ * "started" daje klientowi `name` od razu (gdy provider go zna), żeby UI
+ * mogło pokazać "Przygotowuję: {name}..." przed completion args.
+ */
+export type StreamChunkType =
+  | "text_delta"
+  | "tool_call_started"
+  | "tool_call_arguments_delta"
+  | "tool_call_arguments_complete"
+  | "usage_update";
 
 export interface StreamChunk {
-  type: StreamEventType;
-  content?: string;
-  toolCall?: ToolCall;
-  toolResult?: ToolResult;
+  type: StreamChunkType;
+  /** text_delta */
+  textDelta?: string;
+  /** tool_call_* — stable id z provider'a (np. OpenAI call_id). */
+  toolCallId?: string;
+  /** tool_call_started — nazwa funkcji, jeśli provider ją zna od razu. */
+  toolCallName?: string;
+  /** tool_call_arguments_delta — surowy fragment JSON arguments. */
+  argumentsDelta?: string;
+  /** tool_call_arguments_complete — sparsed pełne arguments. */
+  arguments?: Record<string, unknown>;
+  /** usage_update */
   usage?: TokenUsage;
-  finishReason?: FinishReason;
 }
 
 // ─── LLM Provider ────────────────────────────────────────────────
@@ -181,10 +199,19 @@ export interface LLMProvider {
 
 // ─── Chat Stream (SSE events for chat streaming) ────────────────
 
+/**
+ * SSE event types pomiędzy listenerem (serwer) a `chat-component.tsx` (klient).
+ *
+ * Każdy event niesie monotonicznie rosnące `seq` per session — klient trzyma
+ * `lastSeq` i odrzuca eventy które już zaaplikował (deduplication przy
+ * replay buffer / SSE reconnect).
+ */
 export type ChatStreamEventType =
-  | "content_delta"
-  | "server_tool_start"
-  | "server_tool_result"
+  | "text_delta"
+  | "tool_call_pending"
+  | "tool_call_arguments_delta"
+  | "tool_call_arguments_complete"
+  | "tool_call_executed"
   | "interactive_tool_request"
   | "usage_update"
   | "done"
@@ -193,13 +220,36 @@ export type ChatStreamEventType =
 export interface ChatStreamEvent {
   type: ChatStreamEventType;
   sessionId: string;
-  content?: string;
-  toolCall?: ToolCall;
+  /** Monotonicznie rosnący per session — klient dedupuje. */
+  seq: number;
+  /** ID wiadomości asystenta do której event się odnosi. */
+  messageId?: string;
+
+  /** text_delta */
+  textDelta?: string;
+
+  /** tool_call_pending / tool_call_arguments_delta / tool_call_arguments_complete / tool_call_executed */
+  toolCallId?: string;
+  toolCallName?: string;
+  argumentsDelta?: string;
+  arguments?: Record<string, unknown>;
+
+  /** tool_call_executed */
   toolResult?: ToolResult;
+  executionCount?: number;
+
+  /** interactive_tool_request — multi tools awaiting user input */
   toolCalls?: ToolCall[];
+
+  /** usage_update */
   usage?: TokenUsage;
+
+  /** done */
   finishReason?: FinishReason;
-  executionCount?: number;
+  /** done — total liczba seq w tej sesji (klient może sanity-checkować). */
+  lastSeq?: number;
+
+  /** error */
   error?: string;
 }